{"slug": "open-weight-vs-open-source-ai-4-models-4-licenses", "title": "Open Weight vs Open Source AI: 4 Models, 4 Licenses", "summary": "Four major autumn 2026 model releases ship downloadable weights, but only two — DeepSeek V4.1 Flash 1017 under plain MIT and Alibaba's Qwen3.8 repository under Apache 2.0 — carry standard open source licenses, while Mistral Large 4 uses a modified MIT license requiring a separate commercial deal for Model-as-a-Service operators and Moonshot AI's Kimi K3.1 Pro carries a custom license. None of the four models meets the Open Source Initiative's definition, which requires exposing training code and meaningful training-data information alongside the weights, so even the MIT and Apache models fall short on data and code inspection. Alibaba's newer flagship Qwen4-Max is API-only with no confirmed open-weight release, widening the gap between what labs call open and what lawyers would call open.", "body_md": "Four of autumn 2026’s biggest model releases put downloadable weights in developers’ hands, yet only two of them ship under a license the open source world would recognize without an asterisk. Mistral [released the weights of Mistral Large 4](https://mistral.ai/news/mistral-large-4-oss/?_bhlid=d3550870675e96e5d0294a05aa0493a9373c23d1) under a modified MIT license that demands a separate commercial deal from anyone running a Model-as-a-Service business. Moonshot AI’s Kimi K3.1 Pro followed with a custom license of its own. DeepSeek V4.1 Flash 1017 sits under plain MIT, and Alibaba’s Qwen3.8 repository carries Apache 2.0.\n\nThat split is the core of the open weight vs open source debate, and it changes what you can legally build. A startup that embeds a model inside its own product can use any of the four. A GPU cloud that wants to resell tokens cannot use two of them without signing something first. Teams that already sorted [open weight model licenses across six families](https://shattered.io/open-weight-models-license-comparison-6-families-16gb-2026/) will find the autumn crop sharper than the summer one, because the biggest checkpoints now carry the strictest terms.\n\nAlibaba complicates the picture further. Qwen3.8 was pitched as the first Qwen-Max-class model whose weights would be opened, but Qwen4-Max, the newer flagship, is offered through the API and no open-weight release of it is confirmed. The gap between what a lab calls open and what a lawyer would call open has never been wider, and it shows up in contracts, hardware bills and token prices.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## Open Weight vs Open Source: Three Tiers, Not Two\n\nOpen source software rests on four freedoms: use the thing for any purpose, study how it works, modify it, and share it with or without changes. Applied to AI, the Open Source Initiative’s framing asks for more than a weights file. A truly open source AI system also exposes the code used to build it and meaningful information about the training data, and the license has to preserve those four freedoms without field-of-use carve-outs. Almost no frontier model clears that bar today, and none of the four models here does on the evidence in their release materials.\n\nOpen weight is the looser label. The lab publishes the trained parameters, you download them, and the license decides what happens next. That license can be a standard text like MIT or Apache 2.0, or it can be something the lab wrote itself. The label tells you nothing about the terms, which is why reading the license line matters more than reading the word open on a launch post. Critics on Hacker News made this exact point about Mistral Large 4, calling it weights-available rather than open in the open source sense, and the 9to6ai write-up of the release agreed that the Open Source Initiative’s definition would side with them.\n\nA practical sort uses three tiers:\n\n- **Open license, open weights.** The weights ship under MIT or Apache 2.0 with no resale or revenue conditions. DeepSeek V4.1 Flash 1017 and the Qwen3.8 repository land here.\n- **Open weights, custom conditions.** You can download, fine-tune and ship the model, but a lab-written clause limits one kind of use. Mistral Large 4 and Kimi K3.1 Pro land here.\n- **Hosted only.** You pay per token and never see the weights. Qwen4-Max, Alibaba’s newest flagship, is the freshest example.\n\nThree questions place any model in a tier. Can you use it commercially without calling the lab? Can you resell inference to third parties without a contract? Can you inspect the training data and the training code? The first answer is yes for all four models once you stay below the revenue thresholds. The second answer splits the field in half. The third answer is no across the board, so even the MIT and Apache models are best described as openly licensed weights and not as fully open source AI systems. DeepSeek does publish a technical report for V4.1 Flash, which helps with study, but a paper is not a dataset.\n\nThe labs themselves are inconsistent about vocabulary. Moonshot calls K3.1 Pro its most capable 3T-class open-weight model. Mistral calls Large 4 open weight. DeepSeek’s V4 preview page says the model is live and open-sourced, a stronger word than the license history of the field supports. Qwen’s GitHub page says the series brings a Qwen-Max-class model to open release. Four labs, four phrasings, and only the license file settles which freedoms you actually hold.\n\n## The Four Models on One Spec Table\n\nRaw specs explain why the license terms matter. Two of these models need a data center to serve, one is small enough to rent a modest GPU node for, and the Qwen family splits across sizes. Maker figures anchor each column, and reported numbers carry a label where only third-party trackers publish them.\n\n| Attribute | Kimi K3.1 Pro | Mistral Large 4 | Qwen3.8 (open family) | DeepSeek V4.1 Flash 1017 | \n|---|---|---|---|---|\n| Key release dates | Announced Oct. 23, weights early Nov. (reported Nov. 4) | Preview Oct. 6, weights Oct. 29 | Qwen3.8-Max Aug. 3, repo README updated Aug. 17 | V4.1 Flash Sept. 10, 1017 checkpoint Oct. 17 | \n| Total parameters | About 2.8T (reported) | 1.05T | 2.4T for Qwen3.8-Max (Alibaba) | 552B | \n| Active parameters | About 104B (reported) | 52B | 95B for Qwen3.8-Max | 8B for input, 16B for output | \n| Architecture | Sparse mixture of experts | Mixture of experts with 1.6B vision encoder | Mixture of experts (Max-class) plus smaller dense variants | Mixture of experts, causal encoder-decoder | \n| Context window | 1,048,576 tokens | 1M tokens | 1M tokens (hosted Max) | 1M tokens | \n| Input modalities | Text and image | Text and image | Text and image, Omni-Flash adds audio and video | Text and image | \n| License name | Kimi K3.1-Pro License (custom) | Mistral Ecosystem Open-Weight License v26.10 | Apache-2.0 | MIT | \n| Resale as hosted service | Separate Moonshot agreement above $20M revenue over 12 months (reported) | Separate Mistral commercial license | Permitted under Apache-2.0 | Permitted under MIT | \n| Attribution at scale | Display requirement above 100M monthly users or $20M monthly revenue (reported) | Keep license notice | Keep license notice | Keep license notice | \n| Weights host | Hugging Face | Hugging Face (FP8, NVFP4, community GGUF) | Hugging Face, ModelScope, GitHub | Hugging Face | \n| Official API price (input / output per 1M) | $2 / $10 | $0.70 / $2.10 | Open release not priced, hosted Max priced separately | $0.30 / $1.20 peak | \n| Successor status | Current Moonshot flagship | Current Mistral flagship | Superseded by closed Qwen4-Max (Nov. 16) | V4.1-Pro not yet launched | \n\nThree patterns jump out. First, the parameter counts diverge by a factor of five between DeepSeek’s 552 billion and Kimi’s roughly 2.8 trillion, while active parameters per token sit in a narrower band of 16 billion to 104 billion. Memory follows total parameters and speed follows active parameters, so Kimi demands the most hardware and DeepSeek the least. Second, every model advertises a one million token window, so context length no longer separates them. Third, the two models with the fewest license strings are also the two with the smallest footprint or the oldest flagship status, which means openness and raw size are pulling in opposite directions.\n\nFor a deeper look at how those parameter counts translate into memory budgets, the [memory math for MoE versus dense open-weight models](https://shattered.io/open-weight-ai-models-moe-vs-dense-280b-vs-27b-memory-math-2026/) walks through the arithmetic, and the [side-by-side spec sheet for open-weight models](https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/) covers the wider field beyond these four.\n\n## License Terms Side by Side\n\nThe license file decides whether a model is a tool or a liability. A legal team asks the same questions of every license, and the answers differ sharply. The grid uses each maker’s published terms and reported summaries where the full text sat behind a repository page.\n\n| Question | Kimi K3.1-Pro License | Mistral Ecosystem Open-Weight License | Apache-2.0 (Qwen3.8) | MIT (DeepSeek V4.1 Flash) | \n|---|---|---|---|---|\n| Use, modify and fine-tune | Yes | Yes | Yes | Yes | \n| Sell the model or derivatives | Yes, keep copyright notice | Yes, keep license notice | Yes | Yes | \n| Run a hosted inference or fine-tuning service | Agreement needed above $20M revenue over any 12 months | Commercial license needed for Model-as-a-Service businesses | Yes | Yes | \n| Embed in a product feature | Yes, not counted as a service | Yes, exempt | Yes | Yes | \n| Internal use only | Yes, exempt | Yes, exempt | Yes | Yes | \n| Brand display at massive scale | Yes, above 100M monthly users or $20M monthly revenue | No | No | No | \n| Explicit patent grant | Not stated in sources reviewed | Not stated in sources reviewed | Yes, standard in Apache-2.0 | No, MIT is silent | \n| Standard, pre-vetted text | No | No, modified MIT | Yes | Yes | \n\nThe bottom two rows explain why compliance teams prefer the standard texts. Apache-2.0 and MIT arrive with decades of legal precedent and pre-approved status in many corporate allow lists. A modified MIT license and a lab-written license each require a fresh review, and the review turns on a phrase that carries real ambiguity: whether a customer has meaningful control over the inputs, parameters or training data. A coding assistant that routes free-form prompts straight into the model looks closer to a service than a narrow summarization feature does. If your product sits near that line, a lawyer should read the clause before launch.\n\n## Model-by-Model Breakdown: What Each License Costs You\n\n### Kimi K3.1 Pro: The Biggest Weights and the Most Strings\n\nMoonshot announced Kimi K3.1 Pro on October 23 as the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a one million token window. A second variant, K3.1 Pro Highspeed, arrived the same day and Moonshot advertises it at two to three times the speed. Weights followed a few days later. Reports put the release on November 4, when Moonshot published the files on Hugging Face under its own account with no access gate, and Enjoy’s teardown of the release counted 1.56 TB spread over 96 files.\n\nThat size is the practical story. The checkpoint is already quantized, with 4-bit weights for its routed experts, and Moonshot offers no smaller build. According to the same teardown, the simplest NVIDIA setup in vLLM is eight data-center GPUs, either one B300 node or two GB300 trays, with an AMD route on one MI355X node, and the serving stack wants vLLM 0.34.0 or later. An eight-GPU B300 system carries about 2.1 TB of GPU memory and draws roughly 14 kW. A laptop is out, and so is a single workstation card.\n\nThe license is the second story. The Kimi K3.1-Pro License lets you use, copy, change, fine-tune and sell the model as long as you keep the copyright notice. Two conditions kick in at scale. A business that hosts the model for third parties and brings in more than $20 million over any 12 months must sign a separate agreement with Moonshot before commercial use. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display the Kimi K3.1-Pro name prominently in its interface. Internal use, embedded features and use through Moonshot’s own products or certified inference partners fall outside both conditions. Enjoy reports that the text matches the Kimi K3 license word for word apart from the model name.\n\nOne more wrinkle matters for security teams. Moonshot says it removed cyber-related data from training for the open-weight release, and the model card reports that its internal vulnerability-research score fell from 90.5% to 81.0% as a result, according to Enjoy. Open weights from this lab therefore come with a deliberately trimmed capability set, which is a form of restriction that no license text captures.\n\n### Mistral Large 4: Modified MIT With One Catch\n\nMistral Large 4, nicknamed Le Chonk in the launch materials, carries 1.05 trillion total parameters, 52 billion active parameters, a 1.6 billion parameter vision encoder and a 1M token window. The public preview opened October 6 and the improved weights landed October 29 on Hugging Face, with FP8 and NVFP4 checkpoints and community GGUF conversions following. Mistral’s docs point to vLLM and SGLang for serving. Large 4 supports three reasoning modes, low, medium and high, and high is the default.\n\nThe license, Mistral Ecosystem Open-Weight License v26.10, starts from MIT and adds one condition. Per the 9to6ai breakdown, anyone may use, copy, modify, fine-tune, distribute and sell the model or derivatives while keeping the license notice. A company that runs a Model-as-a-Service business, meaning it gives third parties inference or post-training access with meaningful control over inputs, parameters or training data, needs a commercial license first. The explicit exemptions cover internal use that never exposes the model to third parties, end-user products where the model sits inside specific features, relaying requests to models hosted by others, and use through Mistral’s own products.\n\nThe trend line deserves attention. Mistral Large 3 shipped under Apache 2.0, according to 9to6ai, so the flagship moved from a standard permissive license to a modified one in a single generation. Vendors with the strongest model of the moment have started pricing the right to resell. Mistral also told the market that it sees potential uplift for real cyberattacks and plans to tighten guardrails, while a more capable cyber variant goes only to vetted partners, starting with Orange Cyberdefense, under contract. The open checkpoint therefore sits one step below a restricted sibling, which echoes the Kimi approach.\n\n### Qwen3.8: Apache 2.0 Weights Next Door to a Closed Flagship\n\nThe QwenLM GitHub repository for Qwen3.8 lists an [Apache-2.0 license](https://github.com/QwenLM/Qwen3.8) and states that Qwen3.8 brings a Qwen-Max-class model to open release for the first time. Alibaba describes Qwen3.8-Max, officially released August 3, as a 2.4 trillion parameter model with 95 billion active parameters. The family also includes Qwen3.8-Flash-Next, announced August 3 as an open-weight model on Hugging Face and ModelScope with managed APIs on QwenCloud, and Qwen3.8-Omni-Flash from September 14, which accepts text, image, audio and video with a 1M token window. A dense 27B member, Qwen3.8-27B, is reported under Apache 2.0 and became the base for many community fine-tunes.\n\nThen the ground shifted. Alibaba’s documented product direction moved to Qwen4-Max, and no open-weight release of it is confirmed. Alibaba’s announcement says its API pricing was reduced relative to Qwen3.8-Max, which is a vendor claim about a hosted model. For license purposes the lesson is simple: a license attaches to a specific checkpoint, not to a brand. The Qwen3.8 weights you downloaded keep the license they shipped with, and the newest flagship stays behind an API. The earlier comparison of [Kimi K3.1 Pro, DeepSeek V4.1 Flash and Qwen4-Max](https://shattered.io/open-weight-models-kimi-k3-1-pro-vs-deepseek-v4-1-flash-qwen4-max-glm-5-3-2026/) covers how that hosted-only flagship stacks up on price.\n\n### DeepSeek V4.1 Flash 1017: MIT and the Smallest Footprint\n\nDeepSeek’s [September 10 announcement](https://www.deepseek.com/en/news/deepseek-v4-1-flash/) presents V4.1 Flash as the smallest model in its new architecture family, a 552 billion parameter mixture of experts with a causal encoder-decoder design that activates 8 billion parameters for input and 16 billion for output. DeepSeek says the KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation, which matters because cache costs dominate many agent workloads. The October 17 build, called 1017, is the current Flash checkpoint behind the deepseek-flash alias. Artificial Analysis lists the model as open weights under the MIT license with a Hugging Face repository, a 1M context window and text and image input.\n\nThis is the cleanest license in the group. MIT carries no resale clause, no revenue threshold and no branding rule, so an inference provider can serve the weights without calling DeepSeek. Artificial Analysis already tracks more than 20 hosting providers for the model, including Baseten, CoreWeave, DeepInfra, DigitalOcean and Fireworks. DeepSeek also retired the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names and said it is phasing out V4-Pro, routing deepseek-v4-pro requests to V4.1 Flash at Flash rates until a V4.1-Pro launches. For a roadmap view across the Chinese labs, the [Tencent Hy4 Apache 2.0 release](https://shattered.io/tencent-hy4-770b-parameters-apache-2-0-open-source-model-2026/) shows the same permissive trend from a fourth lab.\n\n## Benchmarks From Four Sources\n\nNo single leaderboard settles this field, and the numbers here come from four families of evidence: vendor model cards, Artificial Analysis, Vals AI and the SuperIndex aggregator. Treat each column as a data point about a specific test, not as a ranking of overall quality.\n\n| Source and test | Kimi K3.1 Pro | Mistral Large 4 | Qwen3.8 open weights | DeepSeek V4.1 Flash 1017 | \n|---|---|---|---|---|\n| Artificial Analysis Intelligence Index v5 | 46 (reported, v5.0.1) | 26 (reported, v5) | No score found | 33, third of 33 models | \n| Artificial Analysis cost per index task | $3.68 (reported) | $2.06 (reported) | No score found | $0.60 | \n| Artificial Analysis output speed | 32 tok/s (reported) | About 161 tok/s (reported) | No score found | 161 tok/s | \n| Terminal-Bench 4.0, vendor card | 65.7% (Moonshot, Claude Code harness) | 41% (Mistral, +13 points vs preview) | No score found | No vendor figure found | \n| Artificial Analysis Coding Agent Index (Mistral-reported) | Not reported | 48%, tied with GLM-5.3, 47% fewer tokens | No score found | Not reported | \n| Cognition FrontierCode 2.0 (Mistral-reported) | Not reported | 19.7%, leads open weights | No score found | Not reported | \n| Vals AI Vals Index | 62.52%, ninth of 32 (reported) | Not reported | No score found | Not reported | \n| SuperIndex SI score | Not retrieved | Not retrieved | Not retrieved | Not listed | \n| Surge Code blind human preference (Mistral-reported) | Lost to Large 4 (K3 tested) | Preferred over GLM-5.3, K3 and DeepSeek V4.1 Flash | Not tested | Lost to Large 4 | \n\nRead the table in layers. On [Artificial Analysis](https://artificialanalysis.ai/models/mistral-large-4?intelligence-category=text-only-vs-multimodal), Kimi K3.1 Pro leads the open-weight pack at 46, with DeepSeek V4.1 Flash 1017 at 33, GLM-5.3 at 32 and Mistral Large 4 at 26, against a median of 23 for large open-weight models. Artificial Analysis swapped to a new version of its index on October 27, so these scores sit on a different scale from the launch-week numbers. Kimi scored 55 on version 4.3 and 46 on version 5.0.1, and Mistral’s preview scored 38 before the rescale. Never compare scores across index versions.\n\nThe trade-off shows up in cost and speed. Kimi tops the index but runs at 32 tokens per second on Moonshot’s own API, with 146 seconds to the first answer token and 37.1 minutes for an index task, per Enjoy’s reading of Artificial Analysis data. DeepSeek finishes a task for $0.60 and streams at 161 tokens per second, though Artificial Analysis flags it as very verbose, with 174 million output tokens on the index against a 104 million median. Mistral sits between them on cost and wins on token efficiency in its own coding evaluations.\n\nThe Terminal-Bench figures show why vendor cards need care. Moonshot’s 65.7% ran in the Claude Code harness at maximum reasoning effort, while Mistral’s 41% comes from the Artificial Analysis coding agent harness. Different harnesses, different scores, and no head-to-head run exists in the sources reviewed. Readers who want the wider field can compare these results with the [open source LLM rankings that split across four labs](https://shattered.io/open-source-llm-leaderboards-disagree-kimi-k3-1-pro-glm-5-3-2026/) and the [ranked list of five open models by input price](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/). Qwen has no independent score for its open checkpoints in the evidence gathered here, which is itself a finding: the strongest open Qwen release is a generation old and the benchmark attention moved to the closed Qwen4-Max.\n\n## Pricing: API Rates Versus Renting Your Own Hardware\n\nOpen weights do not make a model free. They move the bill from a per-token meter to a hardware lease, and the break-even depends on volume. Official API prices convert into tokens once you fix a budget. The budget of $52,500 equals one month of eight rented B300 GPUs at $8.99 per GPU per hour on Runpod ($71.92 an hour for 730 hours), the figure Enjoy used for Kimi K3.1 Pro. The token columns are derived arithmetic assuming five input tokens per output token with no caching.\n\n| Model and tier | Input per 1M | Cached input per 1M | Output per 1M | $52,500 buys (input / output) | \n|---|---|---|---|---|\n| Kimi K3.1 Pro | $2.00 | $0.20 | $10.00 | 13.1B / 2.6B tokens | \n| Mistral Large 4 | $0.70 | $0.07 | $2.10 | Not listed | \n| DeepSeek V4.1 Flash, peak | $0.30 | $0.006 | $1.20 | Not listed | \n| DeepSeek V4.1 Flash, off-peak | $0.15 | $0.003 | $0.60 | Not listed | \n| Qwen3.8 open weights | No API list price for the open release | Not applicable | Not applicable | Not applicable | \n\nSeveral details change the picture. DeepSeek’s pricing page defines peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays, and off-peak rates are half of peak. Mistral offers batch processing at half price and charges a 10% surcharge for regional inference. Mistral also halved its list price when weights shipped, from a preview rate of $1.36 input and $4.18 output down to the current $0.70 and $2.10. Moonshot’s price of $2 and $10 sits a third below the earlier K3 rate, but it matches the listed prices of Claude Sonnet 5.5 and GPT-6.1 Sol rather than undercutting them, per Enjoy’s comparison.\n\nThe same teardown does the key math. At Kimi’s own API rates, a rented eight-GPU node buys the same tokens you could have bought directly for roughly the node’s monthly rent, before storage, networking and staff. Self-hosting Kimi K3.1 Pro therefore pays off only for steady traffic at that scale or when data cannot leave your network. DeepSeek and Mistral change the equation because the API is so cheap that a rented node rarely beats it on cost alone. The reason to self-host those two is control, residency or fine-tuning, not savings. The [DeepSeek, Qwen and Mistral comparison](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/) goes deeper on API economics.\n\n## Real-World Examples: What People Built With These Weights\n\nLicense fine print turns concrete the moment someone ships something. Seven shipped projects show where each license helped, where it constrained, and where the hardware bill dominated. Several come from the community tracking in Jiufeng’s Qwen timeline and from Hugging Face’s own run reports, so treat the figures as reported by those sources.\n\n- **Mistral Large 4 self-hosted.** The weights are downloadable and self-hostable through sources including Hugging Face, with deployment support referenced for vLLM and SGLang. A model of this size calls for a multi-GPU server, not a single consumer card.\n- **Kimi K3.1 Pro on rented B300s.** Enjoy priced a month of eight B300 GPUs at about $52,500 and found that the same money buys roughly 13.1 billion input and 2.6 billion output tokens from Kimi’s API. The authors also warned that Hugging Face listed re-uploads from other accounts within an hour of the official release, and that vLLM’s launch command includes the –trust-remote-code flag, which executes Python from the repository on your machines.\n- **A legacy-alias migration at DeepSeek.** Teams that called deepseek-v4-pro saw their requests routed to V4.1 Flash at Flash rates from September 14, and the older deepseek-v4-flash names now route to the 1017 build. Anyone who pinned an alias without watching the changelog quietly changed models.\n\nTwo lessons repeat across the list. Permissive licenses breed derivatives faster, as the Arete and DSpine cases show, because nobody has to ask permission or disclose revenue. And the supply chain around open weights needs the same scrutiny as any package registry. The [provenance test from ASPI and the ACSC guides](https://shattered.io/open-weight-ai-models-5g-style-provenance-test-aspi-acsc-two-guides/) argues for exactly that kind of vetting before an organization adopts weights from a public repository.\n\n## What the Quotes Say About the Label\n\nDeveloper Simon Willison, who writes about the Kimi licensing terms on his [Moonshot tag page](https://feeds.simonwillison.net/tags/moonshot/), gave the lab credit for honest vocabulary: **“To our credit, they make no attempt to describe this as an \"open source\" license in their own materials, consistently using the term \"open weight\" in its place.”** That comment addressed the Kimi K3 license, and Enjoy reports that the K3.1-Pro License is the same text with the model name swapped.\n\nThe operative clause appears in the license text Willison quotes. Moonshot AI wrote: **“If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.”** Read it twice. The threshold applies to the whole group’s revenue, and the restriction covers derivative works, so a fine-tune does not escape it.\n\nMistral frames its own terms generously. Its announcement says: **“The model is available to use under the Mistral Ecosystem Open-Weight license, a modified MIT license.”** It then spells out the condition: **“The model is free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially.”** Mistral gets credit for stating the catch in the same breath as the grant. The sentence still describes a license that would fail a strict open source test, because it limits a field of use.\n\nThe Business Times offered the industry-level context when Alibaba launched Qwen4-Max: Chinese tech companies remain a huge force in open-weight AI, and the Qwen series is so far the most downloaded open-weight model family in the world, yet the new flagship shipped without downloadable weights. Put the quotes and the launch together and a pattern emerges. The labs publish weights to win developers, then reserve the newest or most capable checkpoint for their own APIs or their own contracts.\n\n## Use-Case Recommendations: Which License Fits Which Job\n\nThe right pick depends less on benchmark rank than on what your company does with the model. Match the pick to the job, not to the leaderboard.\n\n- **A SaaS startup adding AI features to its own product, under $20 million.** Any of the four works legally. Pick DeepSeek V4.1 Flash 1017 for the lowest cost, since $0.30 and $1.20 per million tokens at peak and half that off-peak is the cheapest official API here, and the MIT text needs no legal review.\n- **An inference provider, GPU cloud or API reseller.** Choose DeepSeek or Qwen. Mistral Large 4 requires a commercial license for Model-as-a-Service businesses, and Kimi K3.1 Pro requires an agreement once group revenue passes $20 million over 12 months. More than 20 hosts already serve DeepSeek’s weights, which shows the market treats MIT as an open door.\n- **A European or regulated team with data-residency rules.** Mistral Large 4 deserves the first test. Mistral says it trained the model from scratch on European infrastructure, and it offers cloud, VPC and on-premises deployment through Mistral Studio, plus availability on Microsoft Foundry. The license permits internal use and embedded features without a deal.\n- **A coding-agent team that wants the strongest open model and has hardware or budget.** Kimi K3.1 Pro leads the open-weight pack on Artificial Analysis at 46, so it earns the slot if you can serve eight B300-class GPUs or pay $2 and $10 per million tokens. Accept the slow output speed and the branding clause if you ever reach 100 million users.\n- **A small team with one or two GPUs.** The big mixture-of-experts models are out of reach. Qwen3.8-27B, licensed Apache 2.0, is the realistic path. Read the[three run paths for open source AI models](https://shattered.io/open-source-ai-models-3-run-paths-mistral-large-4-kimi-k3-1-glm-5-3-2026/) before you buy hardware.\n- **A lab or company that plans to publish its own fine-tune.** Use Qwen3.8 or DeepSeek. Nous Research released both weights and training data for Arete-27B under Apache 2.0, which would be harder to do under a license that treats derivatives as covered by revenue thresholds.\n\nOne extra rule of thumb helps. If your product could plausibly be described as a platform that gives other businesses programmatic access to a model, assume a hosting clause applies and talk to the lab. If your product merely uses the model to power features for your own users, every license here is comfortable. The line blurs for developer tools that expose a raw prompt box, and no sources reviewed settle how the labs interpret it.\n\n## Migration Guide: From a Hosted API to Your Own Open Weights\n\nMoving from a per-token API to self-hosted weights is a legal change before it is a technical one. Work through the steps in order, because skipping the first one is the mistake that costs the most.\n\n1. **Classify your product against the license.** Decide whether you operate a service that gives third parties inference or fine-tuning access, or whether the model powers features inside your own product. Write the answer down and file it next to a copy of the license text.\n2. **Download only from the maker’s own account.** Hugging Face listed re-uploads of Kimi K3.1 Pro within an hour of the official release, and one repository using the model’s name predated the real weights. Verify the organization name and the file hashes before you load anything.\n3. **Size the hardware to the checkpoint.** Kimi K3.1 Pro needs about 1.56 TB of storage and, in vLLM’s simplest recipe, eight data-center GPUs. Mistral Large 4 runs on eight H200s with FP8 or four Blackwell cards with NVFP4. DeepSeek publishes no official hardware requirements in the sources reviewed, so test on a rented node first. Qwen3.8-27B fits one large consumer or workstation GPU in a 4-bit build.\n4. **Pick a runtime that supports your checkpoint.** Kimi’s card recommends vLLM, SGLang and TokenSpeed, and vLLM must be version 0.34.0 or later. Mistral’s open-weight announcement references deployment support for vLLM and SGLang. Treat any runtime flag that executes repository code, such as –trust-remote-code, as a security decision and pin the exact commit.\n5. **Keep the client interface identical.** vLLM and SGLang serve an OpenAI-compatible endpoint, so a base URL change is often the whole client migration. The code block below shows the pattern.\n6. **Run a parity test on your own prompts.** Replay a few hundred real requests against both endpoints and score finished-task quality and cost, not only token price. Reasoning controls can differ between hosted and self-hosted builds, and Mistral Large 4 and DeepSeek both burn extra output tokens at high effort.\n7. **Shift traffic in stages.** Send a small slice first, keep the hosted API as a fallback, and log the exact model identifier returned with each response. DeepSeek’s alias changes show how a name can quietly move to a new checkpoint.\n8. **Re-check the license on every upgrade.** Mistral moved from Apache 2.0 on Large 3 to a modified MIT license on Large 4, and Qwen’s newest flagship left the open tier entirely. Treat each new checkpoint as a new contract.\n\n``` python\nfrom openai import OpenAI\n\nhosted = OpenAI(base_url=\"https://api.your-provider.example/v1\", api_key=\"HOSTED_KEY\")\nlocal = OpenAI(base_url=\"http://10.0.0.12:8000/v1\", api_key=\"not-needed\")\n\ndef ask(client, model, prompt):\n    reply = client.chat.completions.create(\n        model=model,\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n        max_tokens=512,\n    )\n    return reply.choices[0].message.content\n\nprompt = \"Summarize the license obligations in this repository.\"\nprint(ask(hosted, \"deepseek-flash\", prompt))\nprint(ask(local, \"my-local-model\", prompt))\nsha256sum LICENSE > LICENSE.sha256\ngrep -n -i -E \"service|monthly active|revenue|attribution|display\" LICENSE\n```\n\nThe second snippet is a crude screen, not legal review. It surfaces the lines that matter on the two restricted licenses: the hosting condition on Mistral and Moonshot, and the display requirement on Moonshot. On MIT and Apache texts it should return almost nothing, and that near-silence is the point.\n\n## Pros and Cons of Each Model\n\n**Kimi K3.1 Pro.** The pros: highest Artificial Analysis score among open weights at 46, a one million token window, and API pricing a third below the previous K3 generation. The cons: a custom license with a $20 million hosting threshold and a 100 million user branding rule, a 1.56 TB checkpoint that needs about eight top-end GPUs, only 32 tokens per second on Moonshot’s own API, and a deliberately trimmed cybersecurity capability.\n\n**Mistral Large 4.** The pros: Mistral-reported coding leadership among open weights (48% on the Coding Agent Index, 19.7% on FrontierCode 2.0), token efficiency, three reasoning modes, broad runtime support and European training infrastructure. The cons: the hosting restriction, a modified license that broke from Large 3’s Apache 2.0, a 26 on the Artificial Analysis index that trails Kimi and DeepSeek, and a hardware floor of four Blackwell GPUs for comfortable serving.\n\n**Qwen3.8.** The pros: Apache 2.0 with a patent grant, a family that spans a Max-class mixture of experts down to a 27B dense model, and the deepest community of derivatives. The cons: the open generation is now behind the closed Qwen4-Max, no independent benchmark appeared for the open weights in the sources reviewed, and the smaller Qwen4 releases carry no date.\n\n**DeepSeek V4.1 Flash 1017.** The pros: plain MIT, the smallest active footprint at 16 billion parameters for output, a cost of $0.60 per Artificial Analysis task, 161 tokens per second, and more than 20 hosts. The cons: very verbose output at 174 million tokens on the index, a first-answer wait of about 45 seconds on Artificial Analysis, no published hardware guide, and a deepseek-flash alias that already moved from the September build to the October 1017 checkpoint.\n\n## Verdict: Match the License to the Business First\n\nOn the open weight vs open source question, the data gives a clear ranking by freedom. **DeepSeek V4.1 Flash 1017 is the most open model in the group**, with MIT terms, a 552 billion parameter footprint that is the smallest of the big four, and the cheapest official API at $0.30 and $1.20 per million tokens at peak. It scores 33 on Artificial Analysis, third among 33 models, which makes the freedom cost little in quality. Qwen3.8 follows on license terms, because Apache 2.0 is just as clean, though its flagship generation has already left the open tier.\n\n**Kimi K3.1 Pro is the strongest open-weight model and the most conditional.** Its 46 on Artificial Analysis beats DeepSeek by 13 points and Mistral by 20, but the license adds a $20 million hosting trigger and a 100 million user branding rule, and the weights need roughly $52,500 a month in rented B300s to serve. **Mistral Large 4 is the pick for European sovereignty and coding-agent efficiency**, provided you are not reselling inference. Its single license condition is narrow, honest and easy to avoid for most product teams.\n\nTwo successors are announced but not yet available. DeepSeek says V4-Pro requests route to V4.1 Flash until a V4.1-Pro launches, and Alibaba says smaller Qwen4 versions are coming, with no dates for either. Both launches could reshuffle the openness ranking, so check the license file the day a new checkpoint lands. For related coverage, browse the [AI and machine learning section](https://shattered.io/category/ai-machine-learning/). The practical rule holds across every model here: read the license line, not the word open.\n\n## FAQ: Open Weight vs Open Source AI Models\n\n### What is the difference between open weight and open source AI?\n\nOpen weight means the lab publishes the trained parameters for download. Open source AI, in the Open Source Initiative’s framing, also requires the code and meaningful information about training data, under a license that preserves the freedoms to use, study, modify and share. All four models here are open weight, and none fully meets the stricter definition on the evidence in their release materials.\n\n### Is DeepSeek V4.1 Flash open source?\n\nIts weights are MIT-licensed and available on Hugging Face, according to Artificial Analysis, so the license is a genuine open source license. The full training data and pipeline are not published, so it counts as open weight under a permissive license rather than a fully open source AI system.\n\n### Can I build a commercial product on Kimi K3.1 Pro?\n\nYes. The license allows use, modification, fine-tuning and sale. A separate agreement is needed only if you host the model for third parties and your group revenue exceeds $20 million over any 12 months, and a display requirement applies above 100 million monthly users or $20 million in monthly revenue. Features embedded in your own product do not count as hosting, per Enjoy’s reading of the license.\n\n### Does Mistral Large 4 allow me to host it for customers?\n\nNot without a deal if your business is Model-as-a-Service, meaning customers get inference or post-training access with meaningful control over inputs, parameters or training data. Internal use, features embedded in your own product and use through Mistral’s own offerings are exempt.\n\n### Is Qwen4-Max open weight?\n\nNo. Alibaba’s newest confirmed flagship is Qwen4-Max, and no official source confirms open weights for it. The Qwen3.8 repository on GitHub remains under Apache-2.0.\n\n### Which of these models is cheapest to run?\n\nThrough official APIs, DeepSeek V4.1 Flash costs $0.30 input and $1.20 output per million tokens at peak and half that off-peak, against $0.70 and $2.10 for Mistral Large 4 and $2 and $10 for Kimi K3.1 Pro. Self-hosting rarely beats those API rates for the large models unless traffic is steady and heavy, because a rented eight-GPU B300 node costs about $52,500 a month.\n\n### How much hardware do I need to self-host Mistral Large 4 or Kimi K3.1 Pro?\n\nKimi’s 1.56 TB checkpoint needs about eight data-center GPUs in vLLM’s simplest recipe. Mistral Large 4 ran on eight H200s with FP8 at about 65 tokens per second for a single request, on four Blackwell cards with NVFP4, and in 2-bit form on a 512 GB Apple Silicon machine at about 15 tokens per second, according to Hugging Face’s reported runs.\n\n### Do licenses change between model versions?\n\nThey can, and recent history shows it. Mistral Large 3 shipped under Apache 2.0 and Large 4 uses a modified MIT license. A license attaches to each checkpoint, so re-read it whenever you upgrade.", "url": "https://wpnews.pro/news/open-weight-vs-open-source-ai-4-models-4-licenses", "canonical_source": "https://shattered.io/open-weight-vs-open-source-ai-4-models-4-licenses-kimi-mistral-qwen-deepseek-2026/", "published_at": "2026-11-22 07:57:26+00:00", "updated_at": "2026-11-22 08:13:50.209903+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-policy", "ai-startups"], "entities": ["Mistral", "Mistral Large 4", "Moonshot AI", "Kimi K3.1 Pro", "DeepSeek", "DeepSeek V4.1 Flash 1017", "Alibaba", "Qwen3.8"], "also_reported_by": [], "takeaway": null, "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/open-weight-vs-open-source-ai-4-models-4-licenses", "markdown": "https://wpnews.pro/news/open-weight-vs-open-source-ai-4-models-4-licenses.md", "text": "https://wpnews.pro/news/open-weight-vs-open-source-ai-4-models-4-licenses.txt", "jsonld": "https://wpnews.pro/news/open-weight-vs-open-source-ai-4-models-4-licenses.jsonld"}}